X-Git-Url: http://git.meshlink.io/?a=blobdiff_plain;f=utcp.c;h=3a6c506062d70a26171f67c653ecea62a44c168c;hb=9461a0793fb6d6a67345cc4666f78d766856985d;hp=5dc7501a20d805e0d57520196bae31ec41657e78;hpb=275129c9d08b2d29529b2f9d039198168aeb52a7;p=utcp

diff --git a/utcp.c b/utcp.c
index 5dc7501..3a6c506 100644
--- a/utcp.c
+++ b/utcp.c
@@ -54,6 +54,10 @@
 	} while (0)
 #endif
 
+static inline size_t min(size_t a, size_t b) {
+	return a < b ? a : b;
+}
+
 static inline size_t max(size_t a, size_t b) {
 	return a > b ? a : b;
 }
@@ -113,9 +117,14 @@ static void print_packet(struct utcp *utcp, const char *dir, const void *pkt, si
 
 	debug("\n");
 }
+
+static void debug_cwnd(struct utcp_connection *c) {
+	debug("snd.cwnd = %u\n", c->snd.cwnd);
+}
 #else
 #define debug(...) do {} while(0)
 #define print_packet(...) do {} while(0)
+#define debug_cwnd(...) do {} while(0)
 #endif
 
 static void set_state(struct utcp_connection *c, enum state state) {
@@ -382,9 +391,10 @@ static struct utcp_connection *allocate_connection(struct utcp *utcp, uint16_t s
 #endif
 	c->snd.una = c->snd.iss;
 	c->snd.nxt = c->snd.iss + 1;
-	c->rcv.wnd = utcp->mtu;
 	c->snd.last = c->snd.nxt;
-	c->snd.cwnd = utcp->mtu;
+	c->snd.cwnd = (utcp->mtu > 2190 ? 2 : utcp->mtu > 1095 ? 3 : 4) * utcp->mtu;
+	c->snd.ssthresh = ~0;
+	debug_cwnd(c);
 	c->utcp = utcp;
 
 	// Add it to the sorted list of connections
@@ -415,13 +425,13 @@ static void update_rtt(struct utcp_connection *c, uint32_t rtt) {
 	if(!utcp->srtt) {
 		utcp->srtt = rtt;
 		utcp->rttvar = rtt / 2;
-		utcp->rto = rtt + max(2 * rtt, CLOCK_GRANULARITY);
 	} else {
 		utcp->rttvar = (utcp->rttvar * 3 + absdiff(utcp->srtt, rtt)) / 4;
 		utcp->srtt = (utcp->srtt * 7 + rtt) / 8;
-		utcp->rto = utcp->srtt + max(utcp->rttvar, CLOCK_GRANULARITY);
 	}
 
+	utcp->rto = utcp->srtt + max(4 * utcp->rttvar, CLOCK_GRANULARITY);
+
 	if(utcp->rto > MAX_RTO) {
 		utcp->rto = MAX_RTO;
 	}
@@ -468,7 +478,7 @@ struct utcp_connection *utcp_connect_ex(struct utcp *utcp, uint16_t dst, utcp_re
 	pkt.hdr.dst = c->dst;
 	pkt.hdr.seq = c->snd.iss;
 	pkt.hdr.ack = 0;
-	pkt.hdr.wnd = c->rcv.wnd;
+	pkt.hdr.wnd = c->rcvbuf.maxsize;
 	pkt.hdr.ctl = SYN;
 	pkt.hdr.aux = 0x0101;
 	pkt.init[0] = 1;
@@ -507,19 +517,22 @@ void utcp_accept(struct utcp_connection *c, utcp_recv_t recv, void *priv) {
 
 static void ack(struct utcp_connection *c, bool sendatleastone) {
 	int32_t left = seqdiff(c->snd.last, c->snd.nxt);
-	int32_t cwndleft = c->snd.cwnd - seqdiff(c->snd.nxt, c->snd.una);
-	debug("cwndleft = %d\n", cwndleft);
+	int32_t cwndleft = min(c->snd.cwnd, c->snd.wnd) - seqdiff(c->snd.nxt, c->snd.una);
 
 	assert(left >= 0);
 
 	if(cwndleft <= 0) {
-		cwndleft = 0;
-	}
-
-	if(cwndleft < left) {
+		left = 0;
+	} else if(cwndleft < left) {
 		left = cwndleft;
+
+		if(!sendatleastone || cwndleft > c->utcp->mtu) {
+			left -= left % c->utcp->mtu;
+		}
 	}
 
+	debug("cwndleft = %d, left = %d\n", cwndleft, left);
+
 	if(!left && !sendatleastone) {
 		return;
 	}
@@ -538,7 +551,7 @@ static void ack(struct utcp_connection *c, bool sendatleastone) {
 	pkt->hdr.src = c->src;
 	pkt->hdr.dst = c->dst;
 	pkt->hdr.ack = c->rcv.nxt;
-	pkt->hdr.wnd = c->snd.wnd;
+	pkt->hdr.wnd = c->rcvbuf.maxsize;
 	pkt->hdr.ctl = ACK;
 	pkt->hdr.aux = 0;
 
@@ -627,11 +640,19 @@ ssize_t utcp_send(struct utcp_connection *c, const void *data, size_t len) {
 
 	// Add data to send buffer.
 
-	len = buffer_put(&c->sndbuf, data, len);
+	if(is_reliable(c) || (c->state != SYN_SENT && c->state != SYN_RECEIVED)) {
+		len = buffer_put(&c->sndbuf, data, len);
+	} else {
+		return 0;
+	}
 
 	if(len <= 0) {
-		errno = EWOULDBLOCK;
-		return 0;
+		if(is_reliable(c)) {
+			errno = EWOULDBLOCK;
+			return 0;
+		} else {
+			return len;
+		}
 	}
 
 	c->snd.last += len;
@@ -667,6 +688,59 @@ static void swap_ports(struct hdr *hdr) {
 	hdr->dst = tmp;
 }
 
+static void fast_retransmit(struct utcp_connection *c) {
+	if(c->state == CLOSED || c->snd.last == c->snd.una) {
+		debug("fast_retransmit() called but nothing to retransmit!\n");
+		return;
+	}
+
+	struct utcp *utcp = c->utcp;
+
+	struct {
+		struct hdr hdr;
+		uint8_t data[];
+	} *pkt;
+
+	pkt = malloc(sizeof(pkt->hdr) + c->utcp->mtu);
+
+	if(!pkt) {
+		return;
+	}
+
+	pkt->hdr.src = c->src;
+	pkt->hdr.dst = c->dst;
+	pkt->hdr.wnd = c->rcvbuf.maxsize;
+	pkt->hdr.aux = 0;
+
+	switch(c->state) {
+	case ESTABLISHED:
+	case FIN_WAIT_1:
+	case CLOSE_WAIT:
+	case CLOSING:
+	case LAST_ACK:
+		// Send unacked data again.
+		pkt->hdr.seq = c->snd.una;
+		pkt->hdr.ack = c->rcv.nxt;
+		pkt->hdr.ctl = ACK;
+		uint32_t len = min(seqdiff(c->snd.last, c->snd.una), utcp->mtu);
+
+		if(fin_wanted(c, c->snd.una + len)) {
+			len--;
+			pkt->hdr.ctl |= FIN;
+		}
+
+		buffer_copy(&c->sndbuf, pkt->data, 0, len);
+		print_packet(c->utcp, "rtrx", pkt, sizeof(pkt->hdr) + len);
+		utcp->send(utcp, pkt, sizeof(pkt->hdr) + len);
+		break;
+
+	default:
+		break;
+	}
+
+	free(pkt);
+}
+
 static void retransmit(struct utcp_connection *c) {
 	if(c->state == CLOSED || c->snd.last == c->snd.una) {
 		debug("Retransmit() called but nothing to retransmit!\n");
@@ -689,7 +763,7 @@ static void retransmit(struct utcp_connection *c) {
 
 	pkt->hdr.src = c->src;
 	pkt->hdr.dst = c->dst;
-	pkt->hdr.wnd = c->rcv.wnd;
+	pkt->hdr.wnd = c->rcvbuf.maxsize;
 	pkt->hdr.aux = 0;
 
 	switch(c->state) {
@@ -737,7 +811,12 @@ static void retransmit(struct utcp_connection *c) {
 		}
 
 		c->snd.nxt = c->snd.una + len;
-		c->snd.cwnd = utcp->mtu; // reduce cwnd on retransmit
+
+		// RFC 5681 slow start after timeout
+		c->snd.ssthresh = max(c->snd.cwnd / 2, utcp->mtu * 2); // eq. 4
+		c->snd.cwnd = utcp->mtu;
+		debug_cwnd(c);
+
 		buffer_copy(&c->sndbuf, pkt->data, 0, len);
 		print_packet(c->utcp, "rtrx", pkt, sizeof(pkt->hdr) + len);
 		utcp->send(utcp, pkt, sizeof(pkt->hdr) + len);
@@ -1005,6 +1084,8 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 		ptr += 2;
 	}
 
+	bool has_data = len || (hdr.ctl & (SYN | FIN));
+
 	// Try to match the packet to an existing connection
 
 	struct utcp_connection *c = find_connection(utcp, hdr.dst, hdr.src);
@@ -1047,6 +1128,7 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 				c->flags = UTCP_TCP;
 			}
 
+synack:
 			// Return SYN+ACK, go to SYN_RECEIVED state
 			c->snd.wnd = hdr.wnd;
 			c->rcv.irs = hdr.seq;
@@ -1062,7 +1144,7 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 			pkt.hdr.dst = c->dst;
 			pkt.hdr.ack = c->rcv.irs + 1;
 			pkt.hdr.seq = c->snd.iss;
-			pkt.hdr.wnd = c->rcv.wnd;
+			pkt.hdr.wnd = c->rcvbuf.maxsize;
 			pkt.hdr.ctl = SYN | ACK;
 
 			if(init) {
@@ -1099,8 +1181,6 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 
 	// It is for an existing connection.
 
-	uint32_t prevrcvnxt = c->rcv.nxt;
-
 	// 1. Drop invalid packets.
 
 	// 1a. Drop packets that should not happen in our current state.
@@ -1124,41 +1204,43 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 		break;
 	}
 
-	// 1b. Drop packets with a sequence number not in our receive window.
+	// 1b. Discard data that is not in our receive window.
 
-	bool acceptable;
+	if(is_reliable(c)) {
+		bool acceptable;
 
-	if(c->state == SYN_SENT) {
-		acceptable = true;
-	} else if(len == 0) {
-		acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0;
-	} else {
-		int32_t rcv_offset = seqdiff(hdr.seq, c->rcv.nxt);
+		if(c->state == SYN_SENT) {
+			acceptable = true;
+		} else if(len == 0) {
+			acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0;
+		} else {
+			int32_t rcv_offset = seqdiff(hdr.seq, c->rcv.nxt);
 
-		// cut already accepted front overlapping
-		if(rcv_offset < 0) {
-			acceptable = len > (size_t) - rcv_offset;
+			// cut already accepted front overlapping
+			if(rcv_offset < 0) {
+				acceptable = len > (size_t) - rcv_offset;
 
-			if(acceptable) {
-				ptr -= rcv_offset;
-				len += rcv_offset;
-				hdr.seq -= rcv_offset;
+				if(acceptable) {
+					ptr -= rcv_offset;
+					len += rcv_offset;
+					hdr.seq -= rcv_offset;
+				}
+			} else {
+				acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt) + len <= c->rcvbuf.maxsize;
 			}
-		} else {
-			acceptable = seqdiff(hdr.seq, c->rcv.nxt) >= 0 && seqdiff(hdr.seq, c->rcv.nxt) + len <= c->rcvbuf.maxsize;
 		}
-	}
 
-	if(!acceptable) {
-		debug("Packet not acceptable, %u <= %u + %lu < %u\n", c->rcv.nxt, hdr.seq, (unsigned long)len, c->rcv.nxt + c->rcvbuf.maxsize);
+		if(!acceptable) {
+			debug("Packet not acceptable, %u <= %u + %lu < %u\n", c->rcv.nxt, hdr.seq, (unsigned long)len, c->rcv.nxt + c->rcvbuf.maxsize);
 
-		// Ignore unacceptable RST packets.
-		if(hdr.ctl & RST) {
-			return 0;
-		}
+			// Ignore unacceptable RST packets.
+			if(hdr.ctl & RST) {
+				return 0;
+			}
 
-		// Otherwise, continue processing.
-		len = 0;
+			// Otherwise, continue processing.
+			len = 0;
+		}
 	}
 
 	c->snd.wnd = hdr.wnd; // TODO: move below
@@ -1167,6 +1249,12 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 	// ackno should not roll back, and it should also not be bigger than what we ever could have sent
 	// (= snd.una + c->sndbuf.used).
 
+	if(!is_reliable(c)) {
+		if(hdr.ack != c->snd.last && c->state >= ESTABLISHED) {
+			hdr.ack = c->snd.una;
+		}
+	}
+
 	if(hdr.ctl & ACK && (seqdiff(hdr.ack, c->snd.last) > 0 || seqdiff(hdr.ack, c->snd.una) < 0)) {
 		debug("Packet ack seqno out of range, %u <= %u < %u\n", c->snd.una, hdr.ack, c->snd.una + c->sndbuf.used);
 
@@ -1268,7 +1356,6 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 	// 3. Advance snd.una
 
 	advanced = seqdiff(hdr.ack, c->snd.una);
-	prevrcvnxt = c->rcv.nxt;
 
 	if(advanced) {
 		// RTT measurement
@@ -1300,8 +1387,10 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 
 		assert(data_acked >= 0);
 
+#ifndef NDEBUG
 		int32_t bufused = seqdiff(c->snd.last, c->snd.una);
 		assert(data_acked <= bufused);
+#endif
 
 		if(data_acked) {
 			buffer_get(&c->sndbuf, NULL, data_acked);
@@ -1314,13 +1403,27 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 
 		c->snd.una = hdr.ack;
 
-		c->dupack = 0;
-		c->snd.cwnd += utcp->mtu;
+		if(c->dupack) {
+			if(c->dupack >= 3) {
+				c->snd.cwnd = c->snd.ssthresh;
+			}
+
+			c->dupack = 0;
+		}
+
+		// Increase the congestion window according to RFC 5681
+		if(c->snd.cwnd < c->snd.ssthresh) {
+			c->snd.cwnd += min(advanced, utcp->mtu); // eq. 2
+		} else {
+			c->snd.cwnd += max(1, (utcp->mtu * utcp->mtu) / c->snd.cwnd); // eq. 3
+		}
 
 		if(c->snd.cwnd > c->sndbuf.maxsize) {
 			c->snd.cwnd = c->sndbuf.maxsize;
 		}
 
+		debug_cwnd(c);
+
 		// Check if we have sent a FIN that is now ACKed.
 		switch(c->state) {
 		case FIN_WAIT_1:
@@ -1333,7 +1436,7 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 		case CLOSING:
 			if(c->snd.una == c->snd.last) {
 				gettimeofday(&c->conn_timeout, NULL);
-				c->conn_timeout.tv_sec += 60;
+				c->conn_timeout.tv_sec += utcp->timeout;
 				set_state(c, TIME_WAIT);
 			}
 
@@ -1348,12 +1451,26 @@ ssize_t utcp_recv(struct utcp *utcp, const void *data, size_t len) {
 
 			if(c->dupack == 3) {
 				debug("Triplicate ACK\n");
-				//TODO: Resend one packet and go to fast recovery mode. See RFC 6582.
-				//We do a very simple variant here; reset the nxt pointer to the last acknowledged packet from the peer.
-				//Reset the congestion window so we wait for ACKs.
-				c->snd.nxt = c->snd.una;
-				c->snd.cwnd = utcp->mtu;
-				start_retransmit_timer(c);
+
+				// RFC 5681 fast recovery
+				c->snd.ssthresh = max(c->snd.cwnd / 2, utcp->mtu * 2); // eq. 4
+				c->snd.cwnd = max(c->snd.ssthresh + 3 * utcp->mtu, c->sndbuf.maxsize);
+
+				if(c->snd.cwnd > c->sndbuf.maxsize) {
+					c->snd.cwnd = c->sndbuf.maxsize;
+				}
+
+				debug_cwnd(c);
+
+				fast_retransmit(c);
+			} else if(c->dupack > 3) {
+				c->snd.cwnd += utcp->mtu;
+
+				if(c->snd.cwnd > c->sndbuf.maxsize) {
+					c->snd.cwnd = c->sndbuf.maxsize;
+				}
+
+				debug_cwnd(c);
 			}
 		}
 	}
@@ -1397,6 +1514,9 @@ skip_ack:
 			break;
 
 		case SYN_RECEIVED:
+			// This is a retransmit of a SYN, send back the SYNACK.
+			goto synack;
+
 		case ESTABLISHED:
 		case FIN_WAIT_1:
 		case FIN_WAIT_2:
@@ -1472,7 +1592,7 @@ skip_ack:
 
 	// 7. Process FIN stuff
 
-	if((hdr.ctl & FIN) && hdr.seq + len == c->rcv.nxt) {
+	if((hdr.ctl & FIN) && (!is_reliable(c) || hdr.seq + len == c->rcv.nxt)) {
 		switch(c->state) {
 		case SYN_SENT:
 		case SYN_RECEIVED:
@@ -1492,7 +1612,7 @@ skip_ack:
 
 		case FIN_WAIT_2:
 			gettimeofday(&c->conn_timeout, NULL);
-			c->conn_timeout.tv_sec += 60;
+			c->conn_timeout.tv_sec += utcp->timeout;
 			set_state(c, TIME_WAIT);
 			break;
 
@@ -1522,12 +1642,15 @@ skip_ack:
 	}
 
 	// Now we send something back if:
-	// - we advanced rcv.nxt (ie, we got some data that needs to be ACKed)
+	// - we received data, so we have to send back an ACK
 	//   -> sendatleastone = true
 	// - or we got an ack, so we should maybe send a bit more data
 	//   -> sendatleastone = false
 
-	ack(c, len || prevrcvnxt != c->rcv.nxt);
+	if(is_reliable(c) || hdr.ctl & SYN || hdr.ctl & FIN) {
+		ack(c, has_data);
+	}
+
 	return 0;
 
 reset:
@@ -1902,8 +2025,14 @@ void utcp_reset_timers(struct utcp *utcp) {
 			continue;
 		}
 
-		c->rtrx_timeout = now;
-		c->conn_timeout = then;
+		if(timerisset(&c->rtrx_timeout)) {
+			c->rtrx_timeout = now;
+		}
+
+		if(timerisset(&c->conn_timeout)) {
+			c->conn_timeout = then;
+		}
+
 		c->rtt_start.tv_sec = 0;
 	}
 
@@ -2054,19 +2183,24 @@ void utcp_expect_data(struct utcp_connection *c, bool expect) {
 }
 
 void utcp_offline(struct utcp *utcp, bool offline) {
+	struct timeval now;
+	gettimeofday(&now, NULL);
+
 	for(int i = 0; i < utcp->nconnections; i++) {
 		struct utcp_connection *c = utcp->connections[i];
 
-		if(!c->reapable) {
-			utcp_expect_data(c, offline);
-
-			// If we are online again, reset the retransmission timers, but keep the connection timeout as it is,
-			// to prevent peers toggling online/offline state frequently from keeping connections alive
-			// if there is no progress in sending actual data.
-			if(!offline) {
-				gettimeofday(&utcp->connections[i]->rtrx_timeout, NULL);
-				utcp->connections[i]->rtt_start.tv_sec = 0;
+		if(c->reapable) {
+			continue;
+		}
+
+		utcp_expect_data(c, offline);
+
+		if(!offline) {
+			if(timerisset(&c->rtrx_timeout)) {
+				c->rtrx_timeout = now;
 			}
+
+			utcp->connections[i]->rtt_start.tv_sec = 0;
 		}
 	}